This year's work has been dominated by Backgammon. Unlike many of our other products, the game engine for this project involved more than one engine. This was because the original engine was a conversion from another product which did not easily fit within our generic architecture and required modifications to integrate it into the structure shared by our other engines. This version was successful, but we knew it would need upgrading to provide a better game engine. That transformation was initiated and re-published in 2022, but we returned to this product again as more work was needed.
A key enhancement in this product was an experimental learning system that shifted the responsibility for tuning away from the developer. The downside of this approach was that the end product became more of a black box, with tuning decisions no longer in the hands of the programmer. The original learning system was labour-intensive to support, so this year we enhanced the system to require less hands-on maintenance. Along with this, we added tools that increased the visibility and security of the tuning process.
Testing these updated versions was also a lengthy process, requiring long round-robin self-play tournaments. The challenge was that this needed to be run on many cores to gather enough results to draw meaningful conclusions. When improvements were marginal, a large number of results were necessary to prove the enhancement. To facilitate this, we added a file-based system that allowed multiple cores to share results by dumping data between rounds (see image above). A separate process could then periodically collect these results to measure the combined outcome. This system is now part of the generic framework and can be utilised by any game engines we develop, providing added long-term value to the work.
Another significant enhancement was the expansion of the statistics provided within Backgammon, along with the addition of built-in tools that allow players to test the integrity of the program. The latter is always a concern, as the program is often accused of cheating, just like all other Backgammon programs. This required significant public relations efforts as we worked to manage the widespread hostile feedback. We took a gradual, systematic approach to this, testing different types of responses and measuring their impact. We also provided a comprehensive on-line set of tests players could run to prove no cheating and explanations about why it may appear to cheat. Over time, this has proven effective, and now most of our top-ranked reviews are 4 and 5 stars, whereas our offline rivals are largely dominated by 1-star reviews.
There were also other new enhancements related to doubling in Backgammon, covered in one of the articles published in this edition.
Again we returned to GDC in San Francisco. This year is was better organised. The previous year suffered from the loss of staff to run the event, after the tail end of Covid, so some things had not worked out well. However this year, unlike last, I did not re-contract Covid! The primary purpose of this is a reality check, re-connecting with people from the industry. The presentations are high quality and always seem to deliver new insights. Its also healthy to be dragged away from your desk to find out what the rest of the industry is doing.
Chess is still our number one product in the download charts and therefore deserves attention. This year, we accidentally discovered a significant bug that had weakened the program, but in a way that was not easily detectable. Finding such issues often relies on a bit of luck. A helpful factor was the extensive tools we have to monitor evaluation decisions, but even then, we might have missed this.
Of course, with many apps in our portfolio, all of them require maintenance. This includes updates mandated by Google and changes to ad processing as AI Factory has been gradually migrating to header bidding.
A final diversion was a trip to an iGGi "unconference," which I initially attended more as a break. However, it proved to be intense as I found myself leading one research project to design an open game development tool and becoming deeply involved in a second group aiming to use an LLM to provide tutoring advice in Chess. The image above is the only one I have, and it actually conveys a much more relaxed atmosphere than was actually the case!
August 2024
